Abhishek Barali and Aayush Paudel test what happens when an extraction schema requires more answers than a document contains. Models often fill the gap with genuine text of the wrong kind.
- Across eight downloadable models and 558 documents, a shown minimum raises over-counting from 34.5% to 55.4%. Enforcing it during generation raises padding to 90.1%.
- An exact source-presence check accepts 46.6% of extra entries under a shown minimum. A copied email sign-off can pass that check while still being a false request.
- Allowing N/A reduces enforced padding to 51.0%. Saying fewer items are fine helps, but also causes models to omit more real answers.
- An exploratory retry test raises Qwen3.5-9B over-counting from 10.0% to 69.0% among the selected replies sent back with a validation error.
Test format compliance, source presence and semantic correctness separately. The released code and run records make this failure useful for testing extraction pipelines.
Caveat: These are author-reported string-list experiments on downloadable models; closed-model APIs and multi-field records remain untested.